From Kifa Plast's factory floor to AI vision systems for industrial quality control

Carl and Mads Fagerlund. Photo: BugiVugi Kommunikation / DI Produktion.

At the family plastics factory, Carl and Mads Fagerlund spent stretches of time sorting faulty plastic products out by hand. The work was monotonous and demanded full concentration. At some point they agreed it had to be possible to do it more intelligently — so they started experimenting with cameras and artificial intelligence to get a computer to spot the faults itself.

That was the start of Deepvis.

Knowing the problem from the factory floor

Neither of us comes from the software industry. We come from production. For several years we worked at Kifa Plast, the family plastics factory, handling quality control of recyclable plastic cups for large brands.

In periods we manually sorted out defective products, where even small scratches, marks or production faults could mean an entire batch had to be checked by hand again.

“So we have felt the consequences up close.”

— Carl Fagerlund, CEO and co-founder

The factory work also showed how hard consistent quality control is when the assessments are made manually, employee to employee.

“We wanted to build the system we ourselves were missing back then. Something fast, simple and intuitive to use for the operators standing at the machines every day, without external specialists having to be involved.”

— Carl Fagerlund

Making quality control simpler

Automated camera-based quality control already exists. The problem is that traditional solutions usually need specialists, extensive algorithm programming and large volumes of training data — with heavy costs both to implement and to run. And when production changes or new fault types appear, the systems often have to be reconfigured by hand.

That complexity is what we set out to remove.

The system works as an extra set of eyes on the production line, with cameras analysing products continuously. Instead of being programmed for each individual fault, it learns the normal patterns in production and reacts to deviations. That makes it flexible in production with recycled plastic — where colours, structures and surfaces vary batch to batch — and in welding, where small variations can be acceptable from part to part.

A bottleneck many are fighting

Large parts of industry are automated, but visual quality control is still done manually by spot check in many places. That means companies often only inspect a small fraction of what leaves production — while pressure rises from documentation requirements, labour shortages, and greater demands for traceability and consistent quality.

Tested in industry

The technology has been tested with Johannes Pedersen Maskinfabrik in Viborg through MADE, Manufacturing Academy of Denmark. The company makes bicycle holders for trains, where thousands of welds have to be checked.

“Our operator gets very tired if he has to inspect 5,000 parts. With more orders the pressure rises, and then it becomes hard to keep up.”

— Falko Thuesen, Purchasing Manager, Johannes Pedersen Maskinfabrik

In the test, the system assessed the welds the same way the company’s welding coordinator did in 90.2% of cases. In several instances it found faults the manual inspection had missed.